AN END-TO-END APPROACH FOR SUPPORTING NAVIGATION OF AUTONOMOUS VEHICLES IN RURAL ENVIRONMENTS

Griffen D. Agnello

doi:10.7273/000007308

The past decade has shown a rising interest in autonomous driving (AD), and the future applications for it continue to expand into new territories. While there has been some work done for rural environments, the majority of the current work has been focused on urban environments. Rural environments present unique challenges to the algorithms onboard an autonomous vehicle (AV) which hinder them from being deployed sooner in these places. All AVs must be equipped to handle a myriad of environmental variables such as lighting, road conditions, emergency traffic maneuvers, and addressing the presence of natural and man-made objects. Urban settings have luxuries that rural settings are not guaranteed to have. They typically have clean, lane-marked, asphalt roads with plenty of traffic lights and traffic street signs to determine what to actions to take. Many cities also have fast and reliable cellular service, enabling “vehicle-to-everything” (V2X) communication between the vehicle and the traffic infrastructure, cloud databases, and other vehicles with V2X technology on the road. On the other hand, in rural conditions, the road may be unpaved or in poor conditions. In scenarios like agricultural settings, there may hardly be a road at all, with several agricultural vehicles being required to drive on dirt and grass. Rural conditions also do not have as much formal traffic infrastructure in order to help guide the flow of traffic. In addition, depending on the location, the AV’s reliability on V2X communication may be hindered by the lack of cellular infrastructure. In order to deploy autonomous vehicles into any terrain, they must be fully equipped edge computing devices that do not require cloud technology. This means that all computation - all the sensing, perception, motion planning, communication, and other algorithms - must be performed in real time onboard the vehicle to ensure safe and reliable driving. This research proposes an end-to-end approach that will enable further development in deploying AVs in rural environments. Current rural AD research have proposed algorithms to utilize segmentation machine learning models in order to navigate rural areas. The issue is that these algorithms require a framework in order to be fully deployed into a real driving scenario. This research aims to address that issue. A prototype case study is shown that demonstrates how segmentation machine learning models can determine where an AV should drive. In this prototype, we utilize a Jetson Nano edge device and optimize the performance of the model with two optimization compilers. The optimized model then performs inference on images of various rural scenarios from a front-facing dash camera view. The prototype is then expanded upon to be used in a full-scale autonomous driving platform known as Autoware®. This distributed computing software package breaks down each component of AVs into separate computation nodes. The nodes for the vehicle’s perception, planning, control, etc. compute information synchronously. In this research, we deploy a YOLOv11 segmentation model into the Autoware perception module. The model then performs real-time inference on a vehicle in a driving simulator called AWSIM. As the vehicle travels in the simulator, the model is able to keep up with the vehicle’s surroundings. This case study is then concluded with how autonomous vehicles are able to use the information gained from the segmentation models to ensure safe driving in rural environments.

AN END-TO-END APPROACH FOR SUPPORTING NAVIGATION OF AUTONOMOUS VEHICLES IN RURAL ENVIRONMENTS

Files and links (1)

Abstract

Metrics

Details